figma-to-weaverse

Figma to Weaverse

Turn a Figma design into maintainable Weaverse pages. This is the Figma input adapter — the sibling of cloning-websites-to-weaverse, but the source is a Figma file instead of a live URL. Both adapters converge on the same downstream skill, generating-weaverse-project-json.

Figma file / frame ─→ figma-to-weaverse ─→ generating-weaverse-project-json ─→ import / content API
   (Figma MCP)        (tokens + manifest + preview)        (project JSON)

The two adapters share the same downstream rules. Section matching, deep schema verification, the content manifest, the clone preview route, and the brand-guideline conflict rule are identical to cloning-websites-to-weaverse — read that skill for the full matching logic. This skill only replaces the extraction layer (Figma MCP instead of Firecrawl) and adds Figma-specific gotchas.

When to Use

A figma.com URL (file, frame, or selection) is the source of a page or storefront
"Implement this design in Weaverse", "build the homepage from Figma"
Multi-frame flows that map to multiple Weaverse pages

Do not use it for live-website cloning (use cloning-websites-to-weaverse) or tiny copy edits.

Required Inputs

Figma URL with a node-id (a page/frame), or the current selection. A file-only URL is not enough — ask for a node link.
.guides/brand-guideline.md
sections.md
app/weaverse/components.ts
Target page type and route (see Preflight)
For delivery: Weaverse project id + WEAVERSE_API_KEY (see weaverse-content-api)

Preflight (before extracting anything)

Each of these cost a full rework loop when skipped:

Page type and route. Ask whether the design becomes the homepage (INDEX), a landing page (CUSTOM, e.g. /disney), or a per-resource template (COLLECTION/PRODUCT for one handle). It decides which sections are allowed (enabledOn) and how data loads. Don't infer it from the nav link.
Right storefront. PUBLIC_STOREFRONT_ID in .env must match the storefront in .shopify/project.json. A mismatched .env renders another brand's sales channel locally, and published collections look missing. Fix with npx shopify hydrogen env pull.
Right brand guideline. Confirm .guides/brand-guideline.md describes this brand (repos are often cloned from a sibling brand). If it doesn't, follow the Figma tokens, say so in the spec, and ask the user to replace the guideline.
Weaverse project state. Read the project's pages (pages <projectId>): empty COLLECTION/PRODUCT templates render header + footer only on those routes — flag it, it is not caused by your page.
Existing notes. Read prior .figma/*.md specs in the repo; header/footer/menu work and token decisions may already exist.

Required Outputs

Same three deliverables as the website cloning skill, plus the live page:

Content manifest — per-section table of real asset URLs (from Figma asset export, later replaced by Shopify CDN URLs), text, links, Shopify refs, media types.
Clone preview route — a $page branch in app/routes/clone-preview.$page.tsx rendering the full design as React + Tailwind, section-marked, brand-adjusted. User must approve this before section decomposition.
Design spec — section mapping, annotation table, schema boundaries, interaction table (every element that reacts, with its evidence), token mapping.
Live Weaverse page — pushed through the Content API and verified on the deployed storefront, with the page JSON kept in .clone/<page>/.
Visual check PASS — .figma/<page>.visual.json + Figma reference screenshots (blocks and interaction states) in .figma/ref/<page>/, and a final scripts/visual_check.mjs report that says Result: PASS (exceptions listed), including every hover and click interaction. The page is not done without it.

Extraction with Figma MCP (MCP-first)

Use the Figma MCP tools. They return structured design data — more precise than scraped HTML for tokens and layout, but weaker on real interaction (a static design has no runtime behavior).

Load figma-design-to-code before get_design_context, and figma-use before any use_figma write call.

get_design_context on the page frame is the primary call. One call returns reference code, a screenshot, asset URLs, the text/color styles used, and component descriptions. If the response is truncated, read the elided middle from the saved output before writing anything. If it asks about Code Connect, ask the user verbatim and follow the answer.
Harvest annotations. Designer notes come back as data-development-annotations (semantics: H1/H2/H3, "not a heading, it's a link") and data-interaction-annotations (menus, link targets, hover behavior). Extract every one into an annotation table in the spec and map each to a requirement — they are easy to lose in a long response. Figma comments are not readable through MCP; ask the user for them if they matter.
Tokens. get_variable_defs when the file uses variables; otherwise use the style list at the end of the get_design_context response. Map them into project.config.theme only for sitewide values; page-specific colors stay as section settings.
Structure. get_metadata only to orient (pick nodes, find frames). Section boundaries are the top-level frames in the design context.
Interaction model — evidence first, in this order. For every element that can react (buttons, links, cards, nav items, tabs, dropdowns, swatches):
1. Component states. If the node is an instance, check the states of its component before anything else. Find the component set (the design context names the component; get_metadata / search_design_system locate the set node), then get_context_for_code_connect on the set lists its variant properties, e.g. State=Default|Hover|Pressed|Active|Disabled. Read Default first, then Hover, then click states (pressed/active/open/selected): get_design_context on each variant node for the exact deltas (colour, underline, border, shadow, scale, swapped icon, revealed child), and get_screenshot → .figma/ref/<page>/<interaction>.png as the state reference. If MCP can't resolve the main component, read it with a read-only use_figma script (getMainComponentAsync(), the set's variantGroupProperties, the instance's prototype reactions such as ON_HOVER → CHANGE_TO); load figma-use first.
2. Annotations. data-interaction-annotations / development annotations on the node or its section (hover menus and their items, link targets, "opens drawer").
3. No state and no note → ask. Layer and section names (Grows Upward — Demo (hover to test), Carousel Demo, Sort Dropdown) only tell you where to look; they are not a behaviour spec. Mark the element UNCONFIRMED in the interaction table and ask the user to confirm the behaviour (batch the questions, at the latest at the approval checkpoint), offering "static" as an option. Build no behaviour for it until answered.
Interaction table columns: element, trigger (hover / click), evidence (variant: <component> State=Hover (<nodeId>) / annotation: <text> / user: <answer, date> / UNCONFIRMED), expected outcome, state reference image. Every confirmed row becomes an interactions entry in the visual spec (step 11).
Assets. The design context lists asset URLs (figma.com/api/mcp/asset/...). They expire after ~7 days. Download the bytes early; inspect them (some are flat color SVGs, decorative masks, or images meant to be shown flipped/cropped).
Global parts. Header, nav, menu dropdowns and footer usually sit inside every frame. Treat them as layout/theme work (theme settings, Shopify menus, layout components), not as page sections.
Placeholders. Designers leave gray boxes, sample product cards with $00.00, stray labels outside the canvas. Mark them MISSING/placeholder in the manifest and map the block to real data (collection products, Instagram feed) instead of reproducing the placeholders.

Plugin fallback

If the MCP can't reach the file or can't export a specific asset (e.g. a flattened group, an unusual export setting), fall back to a Figma agent plugin export and read its output file. MCP is the primary path; use the plugin only to fill a concrete gap, and note in the manifest which assets came from the fallback.

Workflow

Preflight (above). Read .guides/brand-guideline.md; ensure sections.md is current.
Extract the design (steps above). Save the spec + manifest as .figma/<brand>-<page>.md (token table, annotation table, interaction table, content manifest, gaps). While the node ids are at hand, save a Figma screenshot of every block frame to .figma/ref/<page>/<block>.png and of every interaction state variant to .figma/ref/<page>/<interaction>.png (get_screenshot → curl; the URLs are short-lived). These are the visual-check references.
Build the content manifest — same columns and rules as the cloning skill.
Generate the clone preview route at app/routes/clone-preview.$page.tsx — one $page branch per page, React + Tailwind, real Figma assets, {/* === BLOCK NAME === */} markers. Header/footer are not part of it.
User approval checkpoint — STOP and wait. Present the preview URL (the dev server port printed by npm run dev, not a guessed one) + block count, brand overrides, gaps, and the UNCONFIRMED interaction questions. Iterate until the user approves and every interaction row has evidence.
Move assets to Shopify right after approval. Crop responsive variants (desktop/mobile) and flip/rotate where the design does, then upload with weaverse-content-api → upload. Keep the returned media objects in .clone/<page>/shopify-assets.json. Never ship Figma asset URLs.
Decompose & match. Classify every block (media type, composition, layout mechanism, interaction) and match it with the cloning skill's priority + deep structural verification. Also check each candidate's enabledOn against the page type from preflight.
- Parallelise the audit: one read-only scout per candidate section reporting SUPPORTED/GAP per requirement with file:line.
- Then one implementer per section, all under the same contract: every new setting defaults to the current rendering, no edits to the registry/sections.md/shared files outside their task, no build/format runs; the integrator owns registration, docs, lint and build.
- Prefer a small shared block (e.g. a heading + link header) over adding the same header to many sections by hand.
Integrate. Register new sections, update sections.md, run the formatter on changed files only, typecheck and compare against the pre-change error baseline, build the way CI builds.
Generate the page JSON with generating-weaverse-project-json (Shopify media objects, no default values, presets written explicitly), validate it.
Deliver through the Content API (weaverse-content-api): create-page for CUSTOM/template pages (INDEX already exists), then one PATCH that relinks the root main and creates every item. Sitewide values go through theme-update (back up the old values first); menus go through the admin proxy (menuUpdate, back up the old menu). If the page needs code that isn't deployed yet, deploy right after (or tell the user the live page is broken until then).
Verify on the real route with Playwright — loop until PASS. Write .figma/<page>.visual.json: one block per section ([data-wv-id] selector + reference image; mode: "structure" with a note only for data-driven or placeholder blocks) and one interactions entry per confirmed interaction-table row (hover or click, its evidence as source, the expected outcome, the state reference). Run node <this skill>/scripts/visual_check.mjs .figma/<page>.visual.json from the storefront repo. Spec format, setup and the fix playbook: references/visual-check.md.
- Read report.md; for every ❌ open the screenshot, the reference and the .diff.png (or the interaction's problem list), fix one cause (section settings via Content API, or section code), re-run with --only <block-or-interaction>. Repeat until the full run prints Result: PASS on desktop and mobile.
- Never pass by raising thresholds or switching a block to structure to hide a real design difference.
- Escalate to the user with the diff images when a block makes no progress for three iterations or needs a decision/asset you don't have.
- The script covers JS errors, broken images, overflow, H1 count, section order, clipped rounded elements, and every listed hover/click interaction (each in a fresh tab, failing when the expected state already held before the action). Cover data behaviours as click interactions too: a filter or sort click expects a url change, load more expects the new items visible. Still count products in the rendered DOM, not in loader JSON.
- After deploy, re-run the same spec against the deployed URL (--url https://<storefront>/<path>).
Clean up and hand over. Remove the page's preview branch, commit (with the visual spec and references; the generated .figma/visual/ output can stay untracked), push, watch the deploy workflow of this storefront (sibling-brand workflows in the same repo may fail for unrelated reasons). Report the final visual-check result, the accepted exceptions, and what Studio still needs (page SEO, missing collections/links, unpublished resources).

Figma-specific differences from website cloning

Aspect	Website (Firecrawl)	Figma (MCP)
Design tokens	inferred from CSS/computed styles	explicit via `get_variable_defs` — map directly to theme
Section boundaries	HTML section markers, slider/grid classes	top-level frames / auto-layout containers from `get_metadata`
Composition	read from HTML positioning	read from auto-layout direction/wrap/spacing
Assets	image/video URLs in HTML	`download_assets` exports; no real video `src`
Interaction	real runtime behavior in HTML	none at runtime — read component state variants first, then annotations; ask the user when neither exists. Never derive behaviour from layer names
Responsive variants	`<picture>` / breakpoint classes	separate desktop/mobile frames if the designer made them; otherwise one frame

Red Flags

Skipping the approval checkpoint — same rule as website cloning: the user must approve the clone preview before section decomposition.
Inventing interactions — a static Figma frame has no hover, carousel autoplay or scroll behavior unless a component state variant, a prototype reaction, an annotation or the user says so. Layer names like Carousel Demo or hover to test are a pointer, not evidence. Don't classify a frame as a slider just because it shows multiple cards.
Skipping the instance's states — an instance on the canvas shows one variant, usually Default. The Hover/Pressed variants on its component carry the real hover colour, underline or reveal; implementing hover from guesswork when they exist is a rework loop.
Ignoring Figma variables — get_variable_defs is the cleanest token source you'll ever get. Hardcoding colors/sizes when variables exist throws away the main Figma advantage.
Guessing asset URLs — export real assets with download_assets; mark MISSING — [describe] in the manifest when an export fails, don't substitute a placeholder.
One giant bespoke section per frame — decompose into reusable sections; reuse the registry first.
Re-deriving the matching rules — the section-matching priority and deep schema verification live in cloning-websites-to-weaverse. Follow them; don't reinvent a looser version.
Treating mobile/desktop frames as two pages — they're responsive variants of the same section; map to imageMobile/imageDesktop or breakpoint logic, not separate pages.
Leaving the preview route in the repo — delete it after the Weaverse page is verified.
Building before asking the page type — a page built as a COLLECTION template had to be rebuilt as a CUSTOM page; the product grid section also had to change because the collection-only section can't run elsewhere.
Debugging data with the wrong storefront in .env — "collection not found" may just mean the local token belongs to another brand's channel. Check preflight item 2 before blaming publishing.
Letting Figma asset URLs reach Weaverse — they expire within days and some image transforms block them. Upload to Shopify first.
Dropping annotations — designer notes about heading levels, link targets, menus and hover behavior are requirements; keep the annotation table and tick it off at the end.
Checking hover and click by hand only — an unscripted check is skipped on the next iteration. Every confirmed interaction row is an interactions entry in the visual spec, so a regression fails the run.
Reproducing designer placeholders — gray boxes and sample cards are stand-ins for real data, not content.
Declaring success from a stale dev server — after a PATCH, restart npm run dev; the old process can keep serving previous item data under the same ids.
Declaring the page done without a PASS visual check — eyeballing a screenshot missed clipped tile corners that the check reports as no clipped rounded elements ❌. Run the loop to PASS, desktop and mobile.
Loosening the check instead of fixing the page — raising threshold, deleting a block, or marking it structure to make a real difference go away.

Related skills

cloning-websites-to-weaverse — the website input adapter; owns the shared section-matching, manifest, and preview rules this skill reuses.
generating-weaverse-project-json — downstream; turns this skill's spec + manifest + token table into import-ready JSON.
weaverse-content-api — push or update content into the live project after import.
figma-use / figma-generate-design — Figma MCP plugin skills; load figma-use before any use_figma write call.